Caching a DataFrame tells Spark to keep it in memory (or on disk) after the first time it's computed. This avoids recomputing the same transformations every time you trigger an action. Used well, it can dramatically speed up your pipelines. Used carelessly, it can eat all your memory and make things slower.